README.md

---
pretty_name: DataShield Sample-Level Risk Scores
language:
- en
license: other
task_categories:
- text-generation
tags:
- safety
- alignment
- instruction-tuning
- data-filtering
- risk-scoring
- emnlp
size_categories:
- 10K<n<100K
configs:
- config_name: dolly15k
  default: true
  data_files:
  - split: train
    path: data/dolly15k.jsonl
- config_name: alpaca52k
  data_files:
  - split: train
    path: data/alpaca52k.jsonl
---

# DataShield

This dataset releases sample-level risk scores for **DataShield: Uncovering Risky Fine-Tuning Data Across LLMs Through Consensus Subspace Alignment**, accepted to the **EMNLP Main Conference**.

For the method, code, and complete documentation, see the [DataShield GitHub repository](https://github.com/ZJU-LLM-Safety/DataShield).

<p align="center">
  <img src="assets/framework.png" width="900" alt="DataShield framework">
</p>

## Dataset configurations

| Configuration | Source dataset | Rows |
| --- | --- | ---: |
| `dolly15k` | `databricks/databricks-dolly-15k` | 15,011 |
| `alpaca52k` | `tatsu-lab/alpaca` | 51,974 |

```python
from datasets import load_dataset

dolly = load_dataset("killdevil111/DataShield-Sample-Risk", "dolly15k")
alpaca = load_dataset("killdevil111/DataShield-Sample-Risk", "alpaca52k")
```

Each record contains the source instruction-tuning sample and its final DataShield `risk_score`. A higher score indicates a higher estimated fine-tuning risk. The score is intended for ranking and data filtering rather than as a calibrated probability or binary label.

## Citation

```bibtex
@article{wu2026datashield,
  title={DataShield: Uncovering Risky Fine-Tuning Data Across LLMs Through Consensus Subspace Alignment},
  author={Wu, Zefeng and Qi, Weiwei and Chen, Jielong and Zheng, Tianhang and Hong, Di and Lu, Chaochao and He, Liang and Qin, Zhan and Ren, Kui},
  journal={arXiv preprint arXiv:2607.15081},
  year={2026}
}
```